Modeling of Magnetic Core Losses from Global to Local Scales Based on Multimodal Neural Networks

Abstract

This paper presents a comprehensive framework for modeling magnetic core losses, validated through a systematic multi-scale approach. The framework is built on two pillars: a suite of physically-grounded data augmentation and training optimization techniques, and a multimodal neural network that synergizes CNNs and Transformers. Data augmentation techniques—including random period shifts, waveform flipping, noise injection, and parameter perturbations—simulate real-world waveform irregularities, enabling direct applicability to arbitrary-phase and non-periodic inputs. Complementary training strategies ensure stable learning. The core network architecture dynamically fuses scalar parameters and sequential waveforms via attention mechanisms. Extensive validation uses both an in-house experimental platform and the public MagNet database, extending beyond common ferrites to include nanocrystalline and iron powder cores, with data for some materials reaching magnetic saturation. The framework generalizes consistently across these materials, with average prediction errors below 3.4% for most, and handles non-integer periodic waveforms like 3C94 with about 2.44% error. Compared to existing methods, the core contribution lies in the holistic framework integrating expanded experimental data, enhancement techniques tailored to real-world variability, and rigorous multi-material validation, offering a flexible and directly applicable solution for predicting core losses in power electronic devices.